# Read allophant results:
import pickle

result_files = {
    "en": "output/allophant_en_sr16000_ipa.pkl",
    "es": "output/allophant_es_sr16000_ipa.pkl",
}

results = {}
for lang, file in result_files.items():
    with open(file, "rb") as f:
        results[lang] = pickle.load(f)

# Show one example of each language:
for lang, data in results.items():
    print(f"Language: {lang}")
    for wav_file, res in data.items():
        print(f"Wav: {wav_file}")
        print(f"Result: {res}")
        break

Language: en
Wav: input_audios/output/words_wav/en/sr16000/rollback_en1075135_en.Jamaica.wav
Result: ['r', 'o̞ː', 'l', 'b', 'æ', 'k']
Language: es
Wav: input_audios/output/words_wav/es/sr16000/baratija_es023913_es.Castellano.wav
Result: ['b', 'a', 'ɾ', 'a', 't̪', 'i', 'x', 'a']

type(res)

list

# parse results:
import pandas as pd
from pathlib import Path

    
def parse_file_name(file_name):
    """Extract data from file_name 
    e.g. desde_es061900_es.Argentina.wav -> desde (word), es061900 (id), es.Argentina (accent)
    """
    parts = file_name.split('_')
    word = parts[0]
    id_ = parts[1]
    accent = '_'.join(parts[2:]).replace('.wav', '')
    return word, id_, accent



# Make df with cols: lang, accent, word, phone_list, id, wav_file:
def make_df(results):
    data = []
    for lang, lang_data in results.items():
        for wav_file, res in lang_data.items():
            file_name = Path(wav_file).stem
            word, id_, accent = parse_file_name(file_name)
            phones = res
            data.append({
                'lang': lang,
                'accent': accent,
                'word': word,
                'phone_list': phones,
                'id': id_,
                'wav_file': f"{file_name}.wav",
            })
    return pd.DataFrame(data)

df_phones = make_df(results)

df_phones.sample(6)

# df_phones.query("word.str.endswith('pivot') & lang == 'en'").sample(6)
df_phones.query("word.str.endswith('r') & lang == 'es'").sample(6)

print(f"es unique words: {df_phones[df_phones['lang'] == 'es']['word'].nunique()}")
print(f"en unique words: {df_phones[df_phones['lang'] == 'en']['word'].nunique()}")

es unique words: 18795
en unique words: 18422

# Combinations of accents:
pd.set_option('display.max_colwidth', None)  # Show full content of each cell

df_tmp = (
    df_phones
    .groupby(['lang', 'word'])
    .agg({'accent': lambda x: ", ".join(sorted(list(set(x))))})
    .reset_index()
    .groupby(['lang', 'accent'], as_index=False)
    .size()
    .sort_values(by=['lang', 'size'], ascending=[True, False])
)
# if freq < X, accent="others":
df_tmp.loc[df_tmp['size'] < 300, 'accent'] = 'otros'
df_tmp = (
    df_tmp
    .groupby(['lang', 'accent'])
    .agg({'size': 'sum'})
    .reset_index()
    .sort_values(by=['lang', 'size'], ascending=[True, False])
)
df_tmp

import plotnine as p9

df_plot = (
    df_phones
    .groupby(['lang', 'accent'])
    .agg({'word': 'nunique'})
    .reset_index()
    .rename(columns={'word': 'unique_words'})
)
p = (
    p9.ggplot(df_plot, p9.aes(x='accent', y='unique_words'))
    + p9.geom_col(fill="lightblue", color="black", alpha=.7)
    + p9.theme_bw()
    + p9.facet_wrap('~lang', scales='free_x')
    + p9.theme(
        figure_size=(6, 3),
        axis_text_x=p9.element_text(rotation=45, hjust=1),
        legend_position='bottom',
    )
    + p9.labs(x="Acento", y="Palabras únicas")
)
p.show()

# import Levenshtein
from itertools import combinations
# from nltk.metrics.aline import delta, align
from IPython.display import Audio, display, Markdown

def play_audio(df, word, lang, accents: list = None):
    """Play audio for a given word and accent."""
    audio_path = f"output/words_wav/{lang}/sr16000"
    df_filtered = df[(df["word"] == word) & (df["lang"] == lang)]
    if accents is not None:
        df_filtered = df_filtered[df_filtered["accent"].isin(accents)]
    if df_filtered.empty:
        print(f"No audio found for word '{word}' in accents {accents}.")
    else:
        for _, row in df_filtered.iterrows():
            audio_file = f"{audio_path}/{row['wav_file']}"
            word = row["word"]
            phones_str = "".join(row["phone_list"])
            accent = row["accent"]
            display(Markdown(f"**{word}** ({accent}): [{phones_str}]"))
            # Play audio
            display(Audio(audio_file, autoplay=False))            

def my_levenshtein(a, b):
    """Because the original one fails with IPA characters"""
    n, m = len(a), len(b)
    # initialize distance matrix
    dp = [[0] * (m + 1) for _ in range(n + 1)]
    for i in range(n + 1):
        dp[i][0] = i
    for j in range(m + 1):
        dp[0][j] = j
    # fill matrix
    for i in range(1, n + 1):
        for j in range(1, m + 1):
            cost = 0 if a[i - 1] == b[j - 1] else 1
            dp[i][j] = min(
                dp[i - 1][j] + 1,      # deletion
                dp[i][j - 1] + 1,      # insertion
                dp[i - 1][j - 1] + cost  # substitution
            )
    return dp[n][m]

def make_distances_df(df, lang='en'):
    # Create df with cols: word, accents (N cols), phones:
    df_words = (
        df
        .query(f"lang == '{lang}'")
        .pivot_table(
            index=["word", "id"],
            columns="accent",
            values="phone_list",
            aggfunc="first",
        )
        .reset_index()
    )
    # df_words.query("word == 'usa'")
    accent_cols = [col for col in df_words.columns if col not in ['word', 'id']]
    results = []
    for _, row in df_words.iterrows():
        word = row['word']
        wordid = row['id']
        for acc1, acc2 in combinations(accent_cols, 2):
            phones1 = row[acc1]
            phones2 = row[acc2]
            if not (isinstance(phones1, list) and isinstance(phones2, list)):
                l_dist = k_dist = None
            elif len(phones1) == 0 or len(phones2) == 0:
                l_dist = k_dist = None
            else:
                # phone_str_1 = "".join(phones1)
                # phone_str_2 = "".join(phones2)
                # l_dist = Levenshtein.distance(phone_str_1, phone_str_2)
                l_dist = my_levenshtein(phones1, phones2)
                # k_dist = delta(phones1, phones2)  # TODO this fails!
            results.append({
                'word': word,
                'wordid': wordid,
                'accent1': acc1,
                'accent2': acc2,
                'phones1': phones1,
                'phones2': phones2,
                'l_dist': l_dist,
                # 'kondrak_distance': k_dist, # TODO this fails!
                # 'n_phones': TODO, # using phone_list
            })
            # if word == "zucchini":
            # # if wordid == "en100165-55":
            #     print(f"Word: {word}, Accents: {acc1}, {acc2}, Phones: {row[acc1]}, {row[acc2]}, Levenshtein: {l_dist}")
    df_dist = pd.DataFrame(results)
    return df_dist

df_dist_en = make_distances_df(df_phones, lang='en')
df_dist_en["lang"] = 'en'
df_dist_es = make_distances_df(df_phones, lang='es')
df_dist_es["lang"] = 'es'

df_dist = pd.concat([df_dist_en, df_dist_es], ignore_index=True)

df_dist_en.sort_values(by='l_dist', ascending=False)

df_dist_en.query("l_dist == 2").sample(2)

# Add relative distance:
df_dist_clean = df_dist.dropna(subset=['l_dist']).copy()
df_dist_clean["max_len"] = df_dist_clean.apply(lambda x: max(len(x['phones1']), len(x['phones2'])), axis=1)
df_dist_clean["l_dist_norm"] = df_dist_clean["l_dist"] / df_dist_clean["max_len"]

df_tmp = df_dist_clean.query("lang == 'es'").query("l_dist_norm > .5")
df_tmp.sample(20)

play_audio(df_phones, "chusma", "es")

df_tmp = df_dist_clean.query("lang == 'en'").query("l_dist_norm > .5")
df_tmp.sample(10)

play_audio(df_phones, "funky", "en", accents=["en.us.south", "en.Jamaica"])

play_audio(df_phones, "anxious", "en", accents=["en.Jamaica", "en.us.us"])

play_audio(df_phones, "dot", "en")

# Avg rel distance:
avg_rel_dist = df_dist_clean.groupby(["lang", "accent1", "accent2"])["l_dist_norm"].mean().reset_index()
avg_rel_dist = avg_rel_dist.sort_values(by="l_dist_norm", ascending=False)
avg_rel_dist.head(10)

# Make a heatmap of the average relative distance:
import plotnine as p9

def heatmap(df, lang):
    df_plot = df.query(f"lang == '{lang}'").copy()
    df_plot["text"] = df_plot["l_dist_norm"].apply(lambda x: f"{x:.2f}")
    p = (
        p9.ggplot(df_plot, p9.aes(x='accent1', y='accent2', fill='l_dist_norm'))
        + p9.geom_tile(color='white')
        + p9.geom_text(p9.aes(label='text'), size=10, color='white')
        + p9.scale_fill_gradient(name='Avg. Levensh.\nnorm.', low='lightblue', high='darkblue')
        + p9.theme_bw()
        + p9.theme(
            figure_size=(6, 4),
            legend_position='right',
            axis_text_x=p9.element_text(rotation=45, hjust=1, size=10),
            axis_text_y=p9.element_text(size=10),
        )
        + p9.labs(x="", y="")
    )
    p.show()

heatmap(avg_rel_dist, lang='en')
heatmap(avg_rel_dist, lang='es')

import numpy as np
import pandas as pd
import plotnine as p9
from sklearn.linear_model import LogisticRegression
from sklearn.feature_extraction.text import CountVectorizer
from IPython.display import display, Markdown, Audio


def contains_subsequence(data, pattern):
    m, n = len(data), len(pattern)
    if n == 0:
        return True   # empty pattern “always” matches
    if n > m:
        return False  # pattern longer than data can’t match
    # slide a window of length n over data
    for i in range(m - n + 1):
        if data[i:i+n] == pattern:
            return True
    return False

def coef_plot(df_coefs, stats_per_class, top_n=5, ncol=1):
    df_plot = df_coefs.groupby("accent").head(top_n).reset_index(drop=True).copy()
    # Make a label for each accent as: "accent\n(recall, rel_freq)"
    df_plot["accent_label"] = df_plot["accent"].apply(
        lambda x: f"{x}\n(recall: {stats_per_class[x][0]:.2f}, freq: {stats_per_class[x][1]:.2f})"
    )
    nrow = int(np.ceil(len(df_plot["accent"].unique()) / ncol))
    p = (
        p9.ggplot(
            df_plot,
            p9.aes(x="coef", y="reorder(phone+'::'+accent_label, coef)")
        )
        + p9.geom_segment(
            p9.aes(xend=0, yend="reorder(phone+'::'+accent_label, coef)"),
            color="navy", size=1, alpha=0.5
        )
        + p9.geom_point(size=1, color="navy")
        + p9.facet_wrap("accent_label", scales="free_y", ncol=ncol)
        + p9.scale_y_discrete(labels=lambda lst: [l.split("::")[0] for l in lst])
        + p9.theme_bw()
        + p9.theme(
            figure_size=(2 * ncol, 1.5 * nrow),
            axis_text_x=p9.element_text(rotation=45, hjust=1),
            axis_text_y=p9.element_text(size=10),
        )
        + p9.labs(
            y="", x="",
        )
        
    )
    return df_plot, p

def most_distinctive_phones(df, ngram_range=(1,1), min_freq=10):
    """Return a df of logistic regression coefficients for each phone
    for each accent, using n-grams of size ngram_order.
    Also return dict accent: (recall, rel_freq).
    """
    vectorizer = CountVectorizer(
        lowercase=False,
        analyzer='word',
        tokenizer=lambda x: x,
        preprocessor=lambda x: x,
        token_pattern=None,
        ngram_range=ngram_range,
        min_df=min_freq,     # Minimum document frequency
        binary=True,  # Binary counts (presence/absence of n-grams)
    )

    X_vectorized = vectorizer.fit_transform(df["phone_list"].values)
    y = df["accent"].values
    model = LogisticRegression(max_iter=1000, random_state=33)
    model.fit(X_vectorized, y)
    coefs = model.coef_ 
    # print accuracy:
    maj_class_freq = pd.Series(y).value_counts(normalize=True).max()
    print(f"Accuracy: {model.score(X_vectorized, y):.3f} (maj class freq: {maj_class_freq:.3f})")
    # If binary, add first line with negative coefficients:
    if len(model.classes_) == 2:
        coefs = np.vstack([-coefs[0], coefs])
    df_coefs = pd.DataFrame(
        coefs.T, 
        index=vectorizer.get_feature_names_out(),
        columns=model.classes_,
    ).reset_index().rename(columns={"index": "phone"})
    # To long format:
    df_coefs_long = df_coefs.melt(id_vars=["phone"], var_name="accent", value_name="coef")
    df_coefs_long = df_coefs_long.sort_values(by=["accent", "coef"], ascending=[True, False])
    # Make dict accent: (recall, rel_freq)
    pred = model.predict(X_vectorized)
    stats_per_class = {}
    for cl in model.classes_:
        recall = np.mean(pred[y == cl] == cl)
        rel_freq = np.mean(y == cl)
        stats_per_class[cl] = (recall, rel_freq)
    return df_coefs_long, stats_per_class

def show_examples(df, phone_ngram, lang, accent, wavs_path="output/words_wav/en/sr16000", seed=33, n=3):
    """For a given phone ngram (list of phones in order) and an accent,
    show examples of words and their audios.
    """
    df_accent = df[(df["accent"] == accent) & (df["lang"] == lang)].copy()
    df_filtered = df_accent[df_accent["phone_list"].apply(lambda x: contains_subsequence(x, phone_ngram))]
    if df_filtered.empty:
        print(f"No examples found for phone ngram {phone_ngram} in accent {accent}.")
    else:
        df_filtered = df_filtered.sample(n=min(n, len(df_filtered)), random_state=seed)
        for _, row in df_filtered.iterrows():
            word = row["word"]
            phones_str = "".join(row["phone_list"])
            audio_file = f"{wavs_path}/{row['wav_file']}"
            display(Markdown(f"{phone_ngram} as in **{word}** ({accent}): [{phones_str}]"))
            display(Audio(audio_file, autoplay=False))

def run_everything(df, ngram_range=(1,1), lang="es", accents=None, seed=33, ncol=1):
    audio_path = f"output/words_wav/{lang}/sr16000"
    df_filtered = df[df["lang"] == lang].copy()
    if accents is None:
        accents = df_filtered["accent"].unique().tolist()
    df_filtered = df_filtered[df_filtered["accent"].isin(accents)].copy()
    df_coefs, stats_per_class = most_distinctive_phones(df_filtered, ngram_range=ngram_range, min_freq=10)
    # # vec, mod = most_distinctive_phones(df_filtered, ngram_order=ngram_order)
    # return vec, mod
    df_plot, p = coef_plot(df_coefs, stats_per_class, top_n=5, ncol=ncol)
    p.show()
    for _, row in df_plot.iterrows():
        phone_ngram = row["phone"].split(" ")
        accent = row["accent"]
        # print(row)
        # print()
        show_examples(df, phone_ngram, lang, accent, wavs_path=audio_path, seed=seed, n=1)

run_everything(df_phones, ngram_range=(1, 1), lang="es", ncol=3)

Accuracy: 0.481 (maj class freq: 0.333)

run_everything(df_phones, ngram_range=(2, 2), lang="es", ncol=3)

Accuracy: 0.540 (maj class freq: 0.333)

# run_everything(df_phones, ngram_range=(3, 3), lang="es", ncol=3)

run_everything(df_phones, ngram_range=(1, 2), lang="en", ncol=3)
# NOTE if binary, the coeff for the negative class are the negative of the overall coeffs.
# So the left plot displays the largest negative coefficients in absolute value.

Accuracy: 0.333 (maj class freq: 0.120)

run_everything(df_phones, ngram_range=(1, 1), lang="en", ncol=3)
# NOTE if binary, the coeff for the negative class are the negative of the overall coeffs.
# So the left plot displays the largest negative coefficients in absolute value.

Accuracy: 0.229 (maj class freq: 0.120)

def logistic_accuracy(df):
    vectorizer = CountVectorizer(
        lowercase=False,
        analyzer='word',
        tokenizer=lambda x: x,
        preprocessor=lambda x: x,
        token_pattern=None,
        ngram_range=(1, 2),
        min_df=10,
        binary=True,
    )
    X_vectorized = vectorizer.fit_transform(df["phone_list"].values)
    y = df["accent"].values
    model = LogisticRegression(max_iter=1000, random_state=33)
    model.fit(X_vectorized, y)
    acc = model.score(X_vectorized, y)
    n_features = X_vectorized.shape[1]
    return acc, n_features

def pairwise_performance_df(df, lang='en', accents=None):
    df_filtered = df[df["lang"] == lang].copy()
    if accents is None:
        accents = df_filtered["accent"].unique().tolist()
    df_filtered = df_filtered[df_filtered["accent"].isin(accents)].copy()
    accent_pairs = combinations(accents, 2)
    results = []
    for acc1, acc2 in accent_pairs:
        # Keep only rows with the two and words in common:
        df_pair = df_filtered[(df_filtered["accent"] == acc1) | (df_filtered["accent"] == acc2)].copy()
        common_words = df_pair["word"].unique()
        df_pair = df_pair[df_pair["word"].isin(common_words)].copy()
        if df_pair.empty:
            print(f"No common words for accents {acc1} and {acc2}.")
            continue
        acc, n_features = logistic_accuracy(df_pair)
        maj_class_freq = df_pair["accent"].value_counts(normalize=True).max()
        n = len(df_pair)
        results.append({
            'accent1': acc1,
            'accent2': acc2,
            'accuracy': acc,
            'maj_class_freq': maj_class_freq,
            'n': n,
            'n_features': n_features,
        })
    df_results = pd.DataFrame(results)
    return df_results

df_acc_en = pairwise_performance_df(df_phones, lang='en')
df_acc_es = pairwise_performance_df(df_phones, lang='es')

df_acc_en["rel_acc"] = df_acc_en["accuracy"] / df_acc_en["maj_class_freq"]
df_acc_es["rel_acc"] = df_acc_es["accuracy"] / df_acc_es["maj_class_freq"]

def heatmap(df):
    df_plot = df.copy()
    # accents = sorted(df_plot["accent1"].unique())
    # df_plot["accent1"] = pd.Categorical(df_plot["accent1"]) #, categories=accents) #, ordered=True)
    # df_plot["accent2"] = pd.Categorical(df_plot["accent2"]) #, categories=accents) #, ordered=True)
    df_plot["text"] = df_plot.apply(lambda d: f"{d['rel_acc']:.2f}\n({d['accuracy']:.2f})", axis=1)
    # return df_plot
    p = (
        p9.ggplot(df_plot, p9.aes(x='accent2', y='accent1', fill='rel_acc'))
        + p9.geom_tile(color='white')
        + p9.geom_text(p9.aes(label='text'), size=9, color='white')
        + p9.scale_fill_gradient(name='XYZ', low='darkblue', high='lightblue')
        + p9.theme_bw()
        + p9.theme(
            figure_size=(6, 4),
            legend_position='right',
            axis_text_x=p9.element_text(rotation=45, hjust=1, size=10),
            axis_text_y=p9.element_text(size=10),
        )
        + p9.labs(x="", y="")
    )
    p.show()

heatmap(df_acc_en)

df_acc_en

def play_audio(df, word, lang, accents: list = None):
    """Play audio for a given word and accent."""
    audio_path = f"output/words_wav/{lang}/sr16000"
    df_filtered = df[(df["word"] == word) & (df["lang"] == lang)]
    if accents is not None:
        df_filtered = df_filtered[df_filtered["accent"].isin(accents)]
    if df_filtered.empty:
        print(f"No audio found for word '{word}' in accents {accents}.")
    else:
        for _, row in df_filtered.iterrows():
            audio_file = f"{audio_path}/{row['wav_file']}"
            word = row["word"]
            phones_str = "".join(row["phone_list"])
            accent = row["accent"]
            display(Markdown(f"**{word}** ({accent}): [{phones_str}]"))
            # Play audio
            display(Audio(audio_file, autoplay=False))

play_audio(df_phones, "zucchini", "en", accents=["en.uk.general"])

from IPython.display import Audio, display, Markdown
from pathlib import Path

def play_audio(word, accent, path_prefix="output/words_wav/es/sr16000"):
    # find wav file containing word and accent:
    audio_files = list(Path(path_prefix).glob(f"{word}_*{accent}.wav"))
    if not audio_files:
        print(f"No audio files found for {word} with accent {accent}")
        return
    audio_path = audio_files[0]  # Take the first matching file
    try:
        display(Markdown(f"**{word}** ({accent})"))
        display(Audio(audio_path, autoplay=True))
    except FileNotFoundError:
        print(f"Audio file not found: {audio_path}")

play_audio("taberna", "es.Mexico")
# play_audio("caverna", "es.Mexico")
# play_audio("taberna", "es.Argentina")
# play_audio("caverna", "es.Argentina")

	lang	accent	word	phone_list	id	wav_file
17938	en	en.Australian	tony	[t, ɑ, n, i]	en1090440	tony_en1090440_en.Australian.wav
98634	en	en.uk.Yorkshire	diversification	[d, æʌ̯, v, ə, s, ɪ, f, ɪ, k, e, ʃ, ə, n]	en1026122-55	diversification_en1026122-55_en.uk.Yorkshire.wav
42831	en	en.us.south	fifth	[p, eː, θ]	en033792	fifth_en033792_en.us.south.wav
192146	es	es.Mexico	falencia	[f, ə, l, ɛ, n, s̪̻, i, a]	es087739	falencia_es087739_es.Mexico.wav
29690	en	en.Jamaica	splurge	[s, p, l, e̞, d̠ʒ]	en082603	splurge_en082603_en.Jamaica.wav
116131	en	en.scot	overcharge	[o, v, ɚː, t̠ʃ, ɑ, ɹ, d̠ʒ]	en1061859	overcharge_en1061859_en.scot.wav

	lang	accent	word	phone_list	id	wav_file
185478	es	es.Mexico	catalizador	[e̞, t̪, a, l, ɪ, z̪, a, ð, o, ɾ]	es036184	catalizador_es036184_es.Mexico.wav
202354	es	es.Argentina	denegar	[d̪, e, n, i, ɣ, ɑ, r]	es059261	denegar_es059261_es.Argentina.wav
181725	es	es.Mexico	interlocutor	[ɪ, n, t̪, e, r, l, o, k, ʊ, t̪, o, ɾ]	es112375	interlocutor_es112375_es.Mexico.wav
175381	es	es.Mexico	informar	[ɪ, n, f, o, r, m, aː]	es109420	informar_es109420_es.Mexico.wav
205125	es	es.Castellano	reconfortar	[r, e, k, o, m, f, ɔ, ɾ, t̪, a]	es164042	reconfortar_es164042_es.Castellano.wav
207569	es	es.Castellano	amasar	[a, m, a, s, a]	es010955	amasar_es010955_es.Castellano.wav

	word	wordid	accent1	accent2	phones1	phones2	l_dist	lang
429410	faq	en1032377	en.Australian	en.scot	[e, f, æʌ̯, k̟, iː]	[f, ɹ, i, k, w, ə, n, d, l̪, ɨ, ø̞, s, k, w, ɛ, s, t̠ʃ, ɨ, n]	19.0	en
429431	faq	en1032377	en.scot	en.uk.rp	[f, ɹ, i, k, w, ə, n, d, l̪, ɨ, ø̞, s, k, w, ɛ, s, t̠ʃ, ɨ, n]	[ɛu, θ, e, k̟, j, uː]	19.0	en
429423	faq	en1032377	en.Jamaica	en.scot	[ɛ, f, e, k̟, j, uː]	[f, ɹ, i, k, w, ə, n, d, l̪, ɨ, ø̞, s, k, w, ɛ, s, t̠ʃ, ɨ, n]	19.0	en
429432	faq	en1032377	en.scot	en.us.south	[f, ɹ, i, k, w, ə, n, d, l̪, ɨ, ø̞, s, k, w, ɛ, s, t̠ʃ, ɨ, n]	[ɛ, f, æʌ̯, i, k̟, iu̜, ʊ]	18.0	en
429430	faq	en1032377	en.scot	en.uk.general	[f, ɹ, i, k, w, ə, n, d, l̪, ɨ, ø̞, s, k, w, ɛ, s, t̠ʃ, ɨ, n]	[ɛu, f, æʌ̯, k, uː]	18.0	en
...	...	...	...	...	...	...	...	...
1326379	zucchini	en100165-55	en.uk.general	en.us.south	NaN	NaN	NaN	en
1326380	zucchini	en100165-55	en.uk.general	en.us.us	NaN	NaN	NaN	en
1326381	zucchini	en100165-55	en.uk.rp	en.us.south	NaN	NaN	NaN	en
1326382	zucchini	en100165-55	en.uk.rp	en.us.us	NaN	NaN	NaN	en
1326383	zucchini	en100165-55	en.us.south	en.us.us	NaN	NaN	NaN	en

	word	wordid	accent1	accent2	phones1	phones2	l_dist	lang	max_len	l_dist_norm
1366106	pareja	es143329	es.Castellano	es.Mexico	[p, ə, ɾ, i, x, a̟]	[p, a, r, e, x, a]	4.0	es	6	0.666667
1375712	sheriff	es177727	es.Castellano	es.Mexico	[ʂ, i, ɛu, ɗ, iː, f]	[ʃ, ɛ, r, iː, f]	4.0	es	6	0.666667
1380907	vaya	es197027	es.Argentina	es.Mexico	[v, ɑ, zʲ, a]	[ɓ, a, ɟʝ, a]	3.0	es	4	0.750000
1380050	tumbar	es194173	es.Castellano	es.Mexico	[ʊ̃, m, β, ɑ]	[ʊ, m, b, a]	3.0	es	4	0.750000
1343604	desdén	es061965	es.Argentina	es.Castellano	[ð, e, v, ɛ, n]	[d̪, ə, s, t̪, e, n]	5.0	es	6	0.833333
1382560	ánimo	es012975	es.Argentina	es.Mexico	[a̟, n, i, m, õ]	[a, n, ɪ, m, o]	3.0	es	5	0.600000
1368286	playa	es151222	es.Argentina	es.Mexico	[l, a̟, ʒʲ, a̟]	[l, a, ɟʝ, a]	3.0	es	4	0.750000
1340858	correa	es052027	es.Castellano	es.Mexico	[o, r, i, a]	[k, ɔ, r, e, a]	3.0	es	5	0.600000
1326952	abrigar	es001470	es.Argentina	es.Mexico	[a, β, ɾ, ɪ, ɣ, a, r]	[a̟, b, r, i, ɡ, a]	6.0	es	7	0.857143
1360071	lumbar	es121927	es.Argentina	es.Castellano	[l, ʊ̃, m, ɓ, a]	[l, ʊ, m, b, æ]	3.0	es	5	0.600000
1367797	pieza	es149403	es.Argentina	es.Mexico	[j, ɛ, s̪̻, a]	[p, i, e, s̪̻, a]	3.0	es	5	0.600000
1368999	postín	es153819	es.Argentina	es.Castellano	[β, o, θ, t̪, i, n]	[ɔ, s, t̪, ɪ, n]	4.0	es	6	0.666667
1364760	ofensivo	es138231	es.Argentina	es.Castellano	[o, f, e, n, s̪̻, ɪ, β, o]	[ɔ, f, ɨ, n, s, iː, β, ɵ]	5.0	es	8	0.625000
1345562	diálogo	es066128	es.Castellano	es.Mexico	[ð, ɪ, a̟, l, o, k]	[ð, i, a, l, o, ɣ, o]	4.0	es	7	0.571429
1374053	rodar	es171827	es.Castellano	es.Mexico	[ɾ, r, o, d̪, æ]	[r, o, ð, a]	3.0	es	5	0.600000
1345974	duro	es070632	es.Argentina	es.Castellano	[d̪, ʊ, ɾ, oː]	[d̪, uː, ɗ, o]	3.0	es	4	0.750000
1345259	disgusto	es068139	es.Castellano	es.Mexico	[ð, ɪ, s̪̻, ɣ, ʊ, s, t̪, o]	[ð, ɨ, z̪, ɡ, u, s, t, o]	5.0	es	8	0.625000
1329673	amo	es011481	es.Argentina	es.Mexico	[a̟, m, õ]	[a, m, o]	2.0	es	3	0.666667
1358243	jersey	es115041	es.Castellano	es.Mexico	[ʔ, ɛ, d, au, s, æʌ̯]	[d̠ʒ, e̞, r, s, ɪ]	5.0	es	6	0.833333
1344056	desnudez	es063730	es.Castellano	es.Mexico	[d̪, ə, s, n, ʊ, ð, æʌ̯]	[ð, e, z, n, uː, ð, e, s]	6.0	es	8	0.750000

	word	wordid	accent1	accent2	phones1	phones2	l_dist	lang	max_len	l_dist_norm
492582	general	en037980	en.uk.general	en.uk.rp	[ʃ, æʌ̯, n, ɚː, ɹ, ai̯]	[d̠ʒ, æʌ̯, n, r̪, o]	4.0	en	6	0.666667
636193	jaguar	en1048722	en.Australian	en.Jamaica	[d̠ʒ, æ, ɡ, i, w, ə]	[d̠ʒ, a̟, ɡ, o]	4.0	en	6	0.666667
640451	jingle	en1049035	en.Irish	en.uk.general	[d̠ʒ, ɪ, ŋ, ɠ]	[t̠ʃ, ɪ, ŋ, ɡ, ai̯]	3.0	en	5	0.600000
747466	monumental	en1057441	en.uk.rp	en.us.us	[m, o, nʲ, y, m, e, n, t, o]	[m, a, n, j, ɵ, m, ɛ, n, t, æʌ̯]	6.0	en	10	0.600000
152313	bumblebee	en1011821	en.uk.rp	en.us.south	[e̞, m, ɓ, aɪ, b, i]	[β, ɐ̃, m, ɓ, ɓ, e, j]	5.0	en	7	0.714286
573201	humdrum	en1043095	en.Irish	en.scot	[h, o̞, m, d, r, o̞, m]	[ɒ, m, d̠ʒ, o, m]	5.0	en	7	0.714286
435347	feel	en033379	en.us.south	en.us.us	[θ, e, o]	[f, ɨ, l]	3.0	en	3	1.000000
326761	dingy	en025134	en.scot	en.us.us	[ɜ̃, ŋ, ʒʲ, ɪ]	[d, ɛ, ŋ, d̠ʒ, i]	4.0	en	5	0.800000
869504	pond	en065254	en.uk.general	en.us.us	[a, n, d]	[p, ɑ, n, ɗ]	3.0	en	4	0.750000
1283994	warbler	en1097795	en.Jamaica	en.uk.rp	[w, o̞, b, l, ɚː]	[b, u, b, l, ai̯]	3.0	en	5	0.600000

EDA¶

Phone distances¶

Logistic¶

Español¶

English¶

Pairwise performance¶

Buscador de ejemplos¶

	lang	accent	size
3	en	en.Irish, en.Jamaica, en.scot, en.uk.Yorkshire, en.uk.general, en.uk.rp, en.us.south, en.us.us	7819
1	en	en.Australian, en.Irish, en.Jamaica, en.scot, en.uk.Yorkshire, en.uk.general, en.uk.rp, en.us.south, en.us.us	7437
4	en	en.Irish, en.Jamaica, en.scot, en.uk.Yorkshire, en.uk.general, en.uk.rp, en.us.us	1025
2	en	en.Australian, en.Irish, en.Jamaica, en.scot, en.uk.Yorkshire, en.uk.general, en.uk.rp, en.us.us	936
0	en	en.Australian, en.Irish, en.Jamaica, en.scot, en.uk.Yorkshire, en.uk.general, en.uk.rp, en.us.south	532
5	en	en.Irish, en.Jamaica, en.scot, en.uk.Yorkshire, en.uk.general, en.us.south, en.us.us	346
6	en	otros	327
7	es	es.Argentina, es.Castellano, es.Mexico	18775
8	es	otros	20

	word	wordid	accent1	accent2	phones1	phones2	l_dist	lang
1232022	underachieve	en1093310	en.uk.general	en.uk.rp	[ʌ, n, d, ɚː, ɹ, ɛu, t̠ʃ, ɨ, v]	[ʌ, n, d̪, ɵ, ɹ, ɛu, t̠ʃ, ɨ, v]	2.0	en
268448	credence	en1020523	en.uk.general	en.us.us	[k, ɻ, i, d, ɨ, n, z]	[k, ɻ, i, d, ə, n, ts]	2.0	en

	lang	accent1	accent2	l_dist_norm
20	en	en.scot	en.us.south	0.527315
2	en	en.Australian	en.scot	0.513059
23	en	en.uk.general	en.us.south	0.488045
18	en	en.scot	en.uk.general	0.483563
5	en	en.Australian	en.us.south	0.480799
25	en	en.uk.rp	en.us.south	0.475033
1	en	en.Australian	en.Jamaica	0.474636
21	en	en.scot	en.us.us	0.470016
19	en	en.scot	en.uk.rp	0.469233
16	en	en.Jamaica	en.us.south	0.469062

	accent1	accent2	accuracy	maj_class_freq	n	n_features	rel_acc
0	en.Jamaica	en.uk.Yorkshire	0.734667	0.500149	36833	2168	1.468896
1	en.Jamaica	en.uk.rp	0.717520	0.504632	36484	2132	1.421867
2	en.Jamaica	en.scot	0.729581	0.500081	36828	2277	1.458924
3	en.Jamaica	en.us.us	0.711429	0.510707	36050	2134	1.393026
4	en.Jamaica	en.uk.general	0.748832	0.500081	36828	2156	1.497421
5	en.Jamaica	en.us.south	0.772727	0.528323	34848	2191	1.462604
6	en.Jamaica	en.Irish	0.716356	0.500639	36775	2130	1.430884
7	en.Jamaica	en.Australian	0.801696	0.673113	27352	1964	1.191027
8	en.uk.Yorkshire	en.uk.rp	0.690012	0.504781	36495	2092	1.366953
9	en.uk.Yorkshire	en.scot	0.717962	0.500068	36839	2260	1.435729
10	en.uk.Yorkshire	en.us.us	0.745182	0.510857	36061	2126	1.458691
11	en.uk.Yorkshire	en.uk.general	0.691197	0.500068	36839	2094	1.382206
12	en.uk.Yorkshire	en.us.south	0.784991	0.528472	34859	2187	1.485398
13	en.uk.Yorkshire	en.Irish	0.707470	0.500788	36786	2116	1.412713
14	en.uk.Yorkshire	en.Australian	0.792822	0.673245	27363	1904	1.177614
15	en.uk.rp	en.scot	0.720307	0.504714	36490	2264	1.427160
16	en.uk.rp	en.us.us	0.694612	0.506076	35712	2088	1.372545
17	en.uk.rp	en.uk.general	0.655248	0.504714	36490	2059	1.298257
18	en.uk.rp	en.us.south	0.747899	0.523703	34510	2142	1.428097
19	en.uk.rp	en.Irish	0.696133	0.503993	36437	2094	1.381235
20	en.uk.rp	en.Australian	0.757459	0.669023	27014	1895	1.132186
21	en.scot	en.us.us	0.746228	0.510789	36056	2262	1.460933
22	en.scot	en.uk.general	0.740511	0.500000	36834	2265	1.481023
23	en.scot	en.us.south	0.787514	0.528404	34854	2303	1.490362
24	en.scot	en.Irish	0.729453	0.500720	36781	2240	1.456806
25	en.scot	en.Australian	0.790226	0.673185	27358	2077	1.173861
26	en.us.us	en.uk.general	0.719797	0.510789	36056	2101	1.409187
27	en.us.us	en.us.south	0.715489	0.517637	34076	2114	1.382221
28	en.us.us	en.Irish	0.696720	0.510069	36003	2078	1.365933
29	en.us.us	en.Australian	0.762603	0.663619	26580	1907	1.149158
30	en.uk.general	en.us.south	0.758765	0.528404	34854	2135	1.435956
31	en.uk.general	en.Irish	0.701531	0.500720	36781	2094	1.401043
32	en.uk.general	en.Australian	0.757475	0.673185	27358	1882	1.125210
33	en.us.south	en.Irish	0.741674	0.527686	34801	2122	1.405522
34	en.us.south	en.Australian	0.773229	0.647687	25378	1938	1.193831
35	en.Irish	en.Australian	0.768541	0.672551	27305	1880	1.142725